Exploring Pretraining Possibilities of Crop Classifiaction Models Using Large-Scale Sentinel-1 Datasets

Since 2014, ESA Sentinel missions have been producing an ever-growing amount of data for Earth Observation. This creates the opportunity to monitor changes in high temporal and spatial resolution, however, interpreting this huge quantity of data is challenging. In recent years, the rapid advancement and widespread adoption of applied Artificial Intelligence (AI) methods made it feasible to create deep learning models for specific Earth Observation applications. Combining Sentinel datasets with the appropriate amount of ground truth, robust pre-trained models can be created and applied to produce good-quality thematic maps for different years and large areas. Due to responsibilities related to the European Union’s Common Agriculture Policy (CAP) and the motivation for regional yield estimation, crop classification is one of the most frequently studied remote sensing problems these days. Several papers investigate the possible methods to construct robust and generally functioning models to map the spatial distribution of crops as accurately as possible.

In this study, we present the development of a modular, pre-trained deep learning model designed specifically for crop type mapping. The model is tailored to classify the most prevalent crops in Hungary, including winter and autumn cereals, corn, sunflower, alfalfa, rapeseed, grasslands, and other significant types. For pre-training, we leverage country-wide Sentinel-1 Synthetic Aperture Radar (SAR) data such as Sigma Naught or polarimetric descriptors from H-A-alpha decomposition, collected during the 2021–2024 time period. This dataset comprises annual time series of Sentinel-1 pixels at a spatial resolution of 20 meters. Our approach builds upon prior findings that Sentinel-1-based crop type classification performs comparably to methods using Sentinel-2 optical data. However, Sentinel-1 has the added advantage of producing consistent and regular time series, as it is unaffected by atmospheric conditions such as cloud cover.

The proposed model employs a hybrid methodology integrating self-supervised and fully supervised learning paradigms. This modular architecture allows for seamless integration of task-specific classifiers, which can be fine-tuned using supervised learning to address both current and future classification requirements effectively. The fully supervised component is supported by an extensive ground truth dataset covering over 60% of Hungary’s total land area. This proprietary dataset is derived from the national agricultural subsidy database, providing detailed and accurate annotations. The abundance and quality of labeled data enable the construction of robust, highly generalizable models, ensuring reliable performance across diverse classification tasks. This methodology offers great potential to advance national-scale operational tasks, such as early land cover prediction and annual crop type mapping.